The field of legged robotics is rapidly advancing, with a focus on developing more agile, adaptable, and robust systems. Recent research has emphasized the importance of whole-body control, dynamic obstacle avoidance, and multimodal sensing for quadrupedal and bipedal robots. Notably, the integration of reinforcement learning, model predictive control, and spiking neural networks has enabled significant improvements in locomotion and manipulation capabilities.
One of the key directions in this field is the development of control frameworks that can seamlessly integrate high-level task planning with low-level whole-body control. This has led to the creation of more autonomous and adaptable robots that can navigate complex environments and perform a variety of tasks.
Another area of focus is the development of more efficient and robust methods for learning agile locomotion behaviors. This includes the use of unsupervised skill discovery, curriculum learning, and bi-level optimization to enable robots to acquire a diverse repertoire of skills for overcoming obstacles.
Some noteworthy papers in this area include: REBot, which introduces a control framework for quadrupedal robots to achieve real-time reflexive obstacle avoidance. ODYSSEY, which presents a unified mobile manipulation framework for agile quadruped robots equipped with manipulators, seamlessly integrating high-level task planning with low-level whole-body control. Whole-Body Coordination for Dynamic Object Grasping with Legged Manipulators, which introduces DQ-Bench, a new benchmark for evaluating dynamic grasping, and DQ-Net, a compact teacher-student framework for inferring grasp configurations from limited perceptual cues.